drivers/infiniband/core/multicast.c

   1 /*
   2  * Copyright (c) 2006 Intel Corporation.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/completion.h>
  34 #include <linux/dma-mapping.h>
  35 #include <linux/err.h>
  36 #include <linux/interrupt.h>
  37 #include <linux/bitops.h>
  38 #include <linux/random.h>
  39
  40 #include <rdma/ib_cache.h>
  41 #include "sa.h"
  42
  43 static void mcast_add_one(struct ib_device *device);
  44 static void mcast_remove_one(struct ib_device *device);
  45
  46 static struct ib_client mcast_client = {
  47         .name   = "ib_multicast",
  48         .add    = mcast_add_one,
  49         .remove = mcast_remove_one
  50 };
  51
  52 static struct ib_sa_client      sa_client;
  53 static struct workqueue_struct  *mcast_wq;
  54 static union ib_gid mgid0;
  55
  56 struct mcast_device;
  57
  58 struct mcast_port {
  59         struct mcast_device     *dev;
  60         spinlock_t              lock;
  61         struct rb_root          table;
  62         atomic_t                refcount;
  63         struct completion       comp;
  64         u8                      port_num;
  65 };
  66
  67 struct mcast_device {
  68         struct ib_device        *device;
  69         struct ib_event_handler event_handler;
  70         int                     start_port;
  71         int                     end_port;
  72         struct mcast_port       port[0];
  73 };
  74
  75 enum mcast_state {
  76         MCAST_IDLE,
  77         MCAST_JOINING,
  78         MCAST_MEMBER,
  79         MCAST_BUSY,
  80         MCAST_ERROR
  81 };
  82
  83 struct mcast_member;
  84
  85 struct mcast_group {
  86         struct ib_sa_mcmember_rec rec;
  87         struct rb_node          node;
  88         struct mcast_port       *port;
  89         spinlock_t              lock;
  90         struct work_struct      work;
  91         struct list_head        pending_list;
  92         struct list_head        active_list;
  93         struct mcast_member     *last_join;
  94         int                     members[3];
  95         atomic_t                refcount;
  96         enum mcast_state        state;
  97         struct ib_sa_query      *query;
  98         int                     query_id;
  99 };
 100
 101 struct mcast_member {
 102         struct ib_sa_multicast  multicast;
 103         struct ib_sa_client     *client;
 104         struct mcast_group      *group;
 105         struct list_head        list;
 106         enum mcast_state        state;
 107         atomic_t                refcount;
 108         struct completion       comp;
 109 };
 110
 111 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 112                          void *context);
 113 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 114                           void *context);
 115
 116 static struct mcast_group *mcast_find(struct mcast_port *port,
 117                                       union ib_gid *mgid)
 118 {
 119         struct rb_node *node = port->table.rb_node;
 120         struct mcast_group *group;
 121         int ret;
 122
 123         while (node) {
 124                 group = rb_entry(node, struct mcast_group, node);
 125                 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
 126                 if (!ret)
 127                         return group;
 128
 129                 if (ret < 0)
 130                         node = node->rb_left;
 131                 else
 132                         node = node->rb_right;
 133         }
 134         return NULL;
 135 }
 136
 137 static struct mcast_group *mcast_insert(struct mcast_port *port,
 138                                         struct mcast_group *group,
 139                                         int allow_duplicates)
 140 {
 141         struct rb_node **link = &port->table.rb_node;
 142         struct rb_node *parent = NULL;
 143         struct mcast_group *cur_group;
 144         int ret;
 145
 146         while (*link) {
 147                 parent = *link;
 148                 cur_group = rb_entry(parent, struct mcast_group, node);
 149
 150                 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
 151                              sizeof group->rec.mgid);
 152                 if (ret < 0)
 153                         link = &(*link)->rb_left;
 154                 else if (ret > 0)
 155                         link = &(*link)->rb_right;
 156                 else if (allow_duplicates)
 157                         link = &(*link)->rb_left;
 158                 else
 159                         return cur_group;
 160         }
 161         rb_link_node(&group->node, parent, link);
 162         rb_insert_color(&group->node, &port->table);
 163         return NULL;
 164 }
 165
 166 static void deref_port(struct mcast_port *port)
 167 {
 168         if (atomic_dec_and_test(&port->refcount))
 169                 complete(&port->comp);
 170 }
 171
 172 static void release_group(struct mcast_group *group)
 173 {
 174         struct mcast_port *port = group->port;
 175         unsigned long flags;
 176
 177         spin_lock_irqsave(&port->lock, flags);
 178         if (atomic_dec_and_test(&group->refcount)) {
 179                 rb_erase(&group->node, &port->table);
 180                 spin_unlock_irqrestore(&port->lock, flags);
 181                 kfree(group);
 182                 deref_port(port);
 183         } else
 184                 spin_unlock_irqrestore(&port->lock, flags);
 185 }
 186
 187 static void deref_member(struct mcast_member *member)
 188 {
 189         if (atomic_dec_and_test(&member->refcount))
 190                 complete(&member->comp);
 191 }
 192
 193 static void queue_join(struct mcast_member *member)
 194 {
 195         struct mcast_group *group = member->group;
 196         unsigned long flags;
 197
 198         spin_lock_irqsave(&group->lock, flags);
 199         list_add(&member->list, &group->pending_list);
 200         if (group->state == MCAST_IDLE) {
 201                 group->state = MCAST_BUSY;
 202                 atomic_inc(&group->refcount);
 203                 queue_work(mcast_wq, &group->work);
 204         }
 205         spin_unlock_irqrestore(&group->lock, flags);
 206 }
 207
 208 /*
 209  * A multicast group has three types of members: full member, non member, and
 210  * send only member.  We need to keep track of the number of members of each
 211  * type based on their join state.  Adjust the number of members the belong to
 212  * the specified join states.
 213  */
 214 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
 215 {
 216         int i;
 217
 218         for (i = 0; i < 3; i++, join_state >>= 1)
 219                 if (join_state & 0x1)
 220                         group->members[i] += inc;
 221 }
 222
 223 /*
 224  * If a multicast group has zero members left for a particular join state, but
 225  * the group is still a member with the SA, we need to leave that join state.
 226  * Determine which join states we still belong to, but that do not have any
 227  * active members.
 228  */
 229 static u8 get_leave_state(struct mcast_group *group)
 230 {
 231         u8 leave_state = 0;
 232         int i;
 233
 234         for (i = 0; i < 3; i++)
 235                 if (!group->members[i])
 236                         leave_state |= (0x1 << i);
 237
 238         return leave_state & group->rec.join_state;
 239 }
 240
 241 static int check_selector(ib_sa_comp_mask comp_mask,
 242                           ib_sa_comp_mask selector_mask,
 243                           ib_sa_comp_mask value_mask,
 244                           u8 selector, u8 src_value, u8 dst_value)
 245 {
 246         int err;
 247
 248         if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
 249                 return 0;
 250
 251         switch (selector) {
 252         case IB_SA_GT:
 253                 err = (src_value <= dst_value);
 254                 break;
 255         case IB_SA_LT:
 256                 err = (src_value >= dst_value);
 257                 break;
 258         case IB_SA_EQ:
 259                 err = (src_value != dst_value);
 260                 break;
 261         default:
 262                 err = 0;
 263                 break;
 264         }
 265
 266         return err;
 267 }
 268
 269 static int cmp_rec(struct ib_sa_mcmember_rec *src,
 270                    struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
 271 {
 272         /* MGID must already match */
 273
 274         if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
 275             memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
 276                 return -EINVAL;
 277         if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
 278                 return -EINVAL;
 279         if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
 280                 return -EINVAL;
 281         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
 282                            IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
 283                            src->mtu, dst->mtu))
 284                 return -EINVAL;
 285         if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
 286             src->traffic_class != dst->traffic_class)
 287                 return -EINVAL;
 288         if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
 289                 return -EINVAL;
 290         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
 291                            IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
 292                            src->rate, dst->rate))
 293                 return -EINVAL;
 294         if (check_selector(comp_mask,
 295                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
 296                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
 297                            dst->packet_life_time_selector,
 298                            src->packet_life_time, dst->packet_life_time))
 299                 return -EINVAL;
 300         if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
 301                 return -EINVAL;
 302         if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
 303             src->flow_label != dst->flow_label)
 304                 return -EINVAL;
 305         if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
 306             src->hop_limit != dst->hop_limit)
 307                 return -EINVAL;
 308         if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
 309                 return -EINVAL;
 310
 311         /* join_state checked separately, proxy_join ignored */
 312
 313         return 0;
 314 }
 315
 316 static int send_join(struct mcast_group *group, struct mcast_member *member)
 317 {
 318         struct mcast_port *port = group->port;
 319         int ret;
 320
 321         group->last_join = member;
 322         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 323                                        port->port_num, IB_MGMT_METHOD_SET,
 324                                        &member->multicast.rec,
 325                                        member->multicast.comp_mask,
 326                                        3000, GFP_KERNEL, join_handler, group,
 327                                        &group->query);
 328         if (ret >= 0) {
 329                 group->query_id = ret;
 330                 ret = 0;
 331         }
 332         return ret;
 333 }
 334
 335 static int send_leave(struct mcast_group *group, u8 leave_state)
 336 {
 337         struct mcast_port *port = group->port;
 338         struct ib_sa_mcmember_rec rec;
 339         int ret;
 340
 341         rec = group->rec;
 342         rec.join_state = leave_state;
 343
 344         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 345                                        port->port_num, IB_SA_METHOD_DELETE, &rec,
 346                                        IB_SA_MCMEMBER_REC_MGID     |
 347                                        IB_SA_MCMEMBER_REC_PORT_GID |
 348                                        IB_SA_MCMEMBER_REC_JOIN_STATE,
 349                                        3000, GFP_KERNEL, leave_handler,
 350                                        group, &group->query);
 351         if (ret >= 0) {
 352                 group->query_id = ret;
 353                 ret = 0;
 354         }
 355         return ret;
 356 }
 357
 358 static void join_group(struct mcast_group *group, struct mcast_member *member,
 359                        u8 join_state)
 360 {
 361         member->state = MCAST_MEMBER;
 362         adjust_membership(group, join_state, 1);
 363         group->rec.join_state |= join_state;
 364         member->multicast.rec = group->rec;
 365         member->multicast.rec.join_state = join_state;
 366         list_move(&member->list, &group->active_list);
 367 }
 368
 369 static int fail_join(struct mcast_group *group, struct mcast_member *member,
 370                      int status)
 371 {
 372         spin_lock_irq(&group->lock);
 373         list_del_init(&member->list);
 374         spin_unlock_irq(&group->lock);
 375         return member->multicast.callback(status, &member->multicast);
 376 }
 377
 378 static void process_group_error(struct mcast_group *group)
 379 {
 380         struct mcast_member *member;
 381         int ret;
 382
 383         spin_lock_irq(&group->lock);
 384         while (!list_empty(&group->active_list)) {
 385                 member = list_entry(group->active_list.next,
 386                                     struct mcast_member, list);
 387                 atomic_inc(&member->refcount);
 388                 list_del_init(&member->list);
 389                 adjust_membership(group, member->multicast.rec.join_state, -1);
 390                 member->state = MCAST_ERROR;
 391                 spin_unlock_irq(&group->lock);
 392
 393                 ret = member->multicast.callback(-ENETRESET,
 394                                                  &member->multicast);
 395                 deref_member(member);
 396                 if (ret)
 397                         ib_sa_free_multicast(&member->multicast);
 398                 spin_lock_irq(&group->lock);
 399         }
 400
 401         group->rec.join_state = 0;
 402         group->state = MCAST_BUSY;
 403         spin_unlock_irq(&group->lock);
 404 }
 405
 406 static void mcast_work_handler(struct work_struct *work)
 407 {
 408         struct mcast_group *group;
 409         struct mcast_member *member;
 410         struct ib_sa_multicast *multicast;
 411         int status, ret;
 412         u8 join_state;
 413
 414         group = container_of(work, typeof(*group), work);
 415 retest:
 416         spin_lock_irq(&group->lock);
 417         while (!list_empty(&group->pending_list) ||
 418                (group->state == MCAST_ERROR)) {
 419
 420                 if (group->state == MCAST_ERROR) {
 421                         spin_unlock_irq(&group->lock);
 422                         process_group_error(group);
 423                         goto retest;
 424                 }
 425
 426                 member = list_entry(group->pending_list.next,
 427                                     struct mcast_member, list);
 428                 multicast = &member->multicast;
 429                 join_state = multicast->rec.join_state;
 430                 atomic_inc(&member->refcount);
 431
 432                 if (join_state == (group->rec.join_state & join_state)) {
 433                         status = cmp_rec(&group->rec, &multicast->rec,
 434                                          multicast->comp_mask);
 435                         if (!status)
 436                                 join_group(group, member, join_state);
 437                         else
 438                                 list_del_init(&member->list);
 439                         spin_unlock_irq(&group->lock);
 440                         ret = multicast->callback(status, multicast);
 441                 } else {
 442                         spin_unlock_irq(&group->lock);
 443                         status = send_join(group, member);
 444                         if (!status) {
 445                                 deref_member(member);
 446                                 return;
 447                         }
 448                         ret = fail_join(group, member, status);
 449                 }
 450
 451                 deref_member(member);
 452                 if (ret)
 453                         ib_sa_free_multicast(&member->multicast);
 454                 spin_lock_irq(&group->lock);
 455         }
 456
 457         join_state = get_leave_state(group);
 458         if (join_state) {
 459                 group->rec.join_state &= ~join_state;
 460                 spin_unlock_irq(&group->lock);
 461                 if (send_leave(group, join_state))
 462                         goto retest;
 463         } else {
 464                 group->state = MCAST_IDLE;
 465                 spin_unlock_irq(&group->lock);
 466                 release_group(group);
 467         }
 468 }
 469
 470 /*
 471  * Fail a join request if it is still active - at the head of the pending queue.
 472  */
 473 static void process_join_error(struct mcast_group *group, int status)
 474 {
 475         struct mcast_member *member;
 476         int ret;
 477
 478         spin_lock_irq(&group->lock);
 479         member = list_entry(group->pending_list.next,
 480                             struct mcast_member, list);
 481         if (group->last_join == member) {
 482                 atomic_inc(&member->refcount);
 483                 list_del_init(&member->list);
 484                 spin_unlock_irq(&group->lock);
 485                 ret = member->multicast.callback(status, &member->multicast);
 486                 deref_member(member);
 487                 if (ret)
 488                         ib_sa_free_multicast(&member->multicast);
 489         } else
 490                 spin_unlock_irq(&group->lock);
 491 }
 492
 493 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 494                          void *context)
 495 {
 496         struct mcast_group *group = context;
 497
 498         if (status)
 499                 process_join_error(group, status);
 500         else {
 501                 spin_lock_irq(&group->port->lock);
 502                 group->rec = *rec;
 503                 if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
 504                         rb_erase(&group->node, &group->port->table);
 505                         mcast_insert(group->port, group, 1);
 506                 }
 507                 spin_unlock_irq(&group->port->lock);
 508         }
 509         mcast_work_handler(&group->work);
 510 }
 511
 512 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 513                           void *context)
 514 {
 515         struct mcast_group *group = context;
 516
 517         mcast_work_handler(&group->work);
 518 }
 519
 520 static struct mcast_group *acquire_group(struct mcast_port *port,
 521                                          union ib_gid *mgid, gfp_t gfp_mask)
 522 {
 523         struct mcast_group *group, *cur_group;
 524         unsigned long flags;
 525         int is_mgid0;
 526
 527         is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
 528         if (!is_mgid0) {
 529                 spin_lock_irqsave(&port->lock, flags);
 530                 group = mcast_find(port, mgid);
 531                 if (group)
 532                         goto found;
 533                 spin_unlock_irqrestore(&port->lock, flags);
 534         }
 535
 536         group = kzalloc(sizeof *group, gfp_mask);
 537         if (!group)
 538                 return NULL;
 539
 540         group->port = port;
 541         group->rec.mgid = *mgid;
 542         INIT_LIST_HEAD(&group->pending_list);
 543         INIT_LIST_HEAD(&group->active_list);
 544         INIT_WORK(&group->work, mcast_work_handler);
 545         spin_lock_init(&group->lock);
 546
 547         spin_lock_irqsave(&port->lock, flags);
 548         cur_group = mcast_insert(port, group, is_mgid0);
 549         if (cur_group) {
 550                 kfree(group);
 551                 group = cur_group;
 552         } else
 553                 atomic_inc(&port->refcount);
 554 found:
 555         atomic_inc(&group->refcount);
 556         spin_unlock_irqrestore(&port->lock, flags);
 557         return group;
 558 }
 559
 560 /*
 561  * We serialize all join requests to a single group to make our lives much
 562  * easier.  Otherwise, two users could try to join the same group
 563  * simultaneously, with different configurations, one could leave while the
 564  * join is in progress, etc., which makes locking around error recovery
 565  * difficult.
 566  */
 567 struct ib_sa_multicast *
 568 ib_sa_join_multicast(struct ib_sa_client *client,
 569                      struct ib_device *device, u8 port_num,
 570                      struct ib_sa_mcmember_rec *rec,
 571                      ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
 572                      int (*callback)(int status,
 573                                      struct ib_sa_multicast *multicast),
 574                      void *context)
 575 {
 576         struct mcast_device *dev;
 577         struct mcast_member *member;
 578         struct ib_sa_multicast *multicast;
 579         int ret;
 580
 581         dev = ib_get_client_data(device, &mcast_client);
 582         if (!dev)
 583                 return ERR_PTR(-ENODEV);
 584
 585         member = kmalloc(sizeof *member, gfp_mask);
 586         if (!member)
 587                 return ERR_PTR(-ENOMEM);
 588
 589         ib_sa_client_get(client);
 590         member->client = client;
 591         member->multicast.rec = *rec;
 592         member->multicast.comp_mask = comp_mask;
 593         member->multicast.callback = callback;
 594         member->multicast.context = context;
 595         init_completion(&member->comp);
 596         atomic_set(&member->refcount, 1);
 597         member->state = MCAST_JOINING;
 598
 599         member->group = acquire_group(&dev->port[port_num - dev->start_port],
 600                                       &rec->mgid, gfp_mask);
 601         if (!member->group) {
 602                 ret = -ENOMEM;
 603                 goto err;
 604         }
 605
 606         /*
 607          * The user will get the multicast structure in their callback.  They
 608          * could then free the multicast structure before we can return from
 609          * this routine.  So we save the pointer to return before queuing
 610          * any callback.
 611          */
 612         multicast = &member->multicast;
 613         queue_join(member);
 614         return multicast;
 615
 616 err:
 617         ib_sa_client_put(client);
 618         kfree(member);
 619         return ERR_PTR(ret);
 620 }
 621 EXPORT_SYMBOL(ib_sa_join_multicast);
 622
 623 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
 624 {
 625         struct mcast_member *member;
 626         struct mcast_group *group;
 627
 628         member = container_of(multicast, struct mcast_member, multicast);
 629         group = member->group;
 630
 631         spin_lock_irq(&group->lock);
 632         if (member->state == MCAST_MEMBER)
 633                 adjust_membership(group, multicast->rec.join_state, -1);
 634
 635         list_del_init(&member->list);
 636
 637         if (group->state == MCAST_IDLE) {
 638                 group->state = MCAST_BUSY;
 639                 spin_unlock_irq(&group->lock);
 640                 /* Continue to hold reference on group until callback */
 641                 queue_work(mcast_wq, &group->work);
 642         } else {
 643                 spin_unlock_irq(&group->lock);
 644                 release_group(group);
 645         }
 646
 647         deref_member(member);
 648         wait_for_completion(&member->comp);
 649         ib_sa_client_put(member->client);
 650         kfree(member);
 651 }
 652 EXPORT_SYMBOL(ib_sa_free_multicast);
 653
 654 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
 655                            union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
 656 {
 657         struct mcast_device *dev;
 658         struct mcast_port *port;
 659         struct mcast_group *group;
 660         unsigned long flags;
 661         int ret = 0;
 662
 663         dev = ib_get_client_data(device, &mcast_client);
 664         if (!dev)
 665                 return -ENODEV;
 666
 667         port = &dev->port[port_num - dev->start_port];
 668         spin_lock_irqsave(&port->lock, flags);
 669         group = mcast_find(port, mgid);
 670         if (group)
 671                 *rec = group->rec;
 672         else
 673                 ret = -EADDRNOTAVAIL;
 674         spin_unlock_irqrestore(&port->lock, flags);
 675
 676         return ret;
 677 }
 678 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
 679
 680 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
 681                              struct ib_sa_mcmember_rec *rec,
 682                              struct ib_ah_attr *ah_attr)
 683 {
 684         int ret;
 685         u16 gid_index;
 686         u8 p;
 687
 688         ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
 689         if (ret)
 690                 return ret;
 691
 692         memset(ah_attr, 0, sizeof *ah_attr);
 693         ah_attr->dlid = be16_to_cpu(rec->mlid);
 694         ah_attr->sl = rec->sl;
 695         ah_attr->port_num = port_num;
 696         ah_attr->static_rate = rec->rate;
 697
 698         ah_attr->ah_flags = IB_AH_GRH;
 699         ah_attr->grh.dgid = rec->mgid;
 700
 701         ah_attr->grh.sgid_index = (u8) gid_index;
 702         ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
 703         ah_attr->grh.hop_limit = rec->hop_limit;
 704         ah_attr->grh.traffic_class = rec->traffic_class;
 705
 706         return 0;
 707 }
 708 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
 709
 710 static void mcast_groups_lost(struct mcast_port *port)
 711 {
 712         struct mcast_group *group;
 713         struct rb_node *node;
 714         unsigned long flags;
 715
 716         spin_lock_irqsave(&port->lock, flags);
 717         for (node = rb_first(&port->table); node; node = rb_next(node)) {
 718                 group = rb_entry(node, struct mcast_group, node);
 719                 spin_lock(&group->lock);
 720                 if (group->state == MCAST_IDLE) {
 721                         atomic_inc(&group->refcount);
 722                         queue_work(mcast_wq, &group->work);
 723                 }
 724                 group->state = MCAST_ERROR;
 725                 spin_unlock(&group->lock);
 726         }
 727         spin_unlock_irqrestore(&port->lock, flags);
 728 }
 729
 730 static void mcast_event_handler(struct ib_event_handler *handler,
 731                                 struct ib_event *event)
 732 {
 733         struct mcast_device *dev;
 734
 735         dev = container_of(handler, struct mcast_device, event_handler);
 736
 737         switch (event->event) {
 738         case IB_EVENT_PORT_ERR:
 739         case IB_EVENT_LID_CHANGE:
 740         case IB_EVENT_SM_CHANGE:
 741         case IB_EVENT_CLIENT_REREGISTER:
 742                 mcast_groups_lost(&dev->port[event->element.port_num -
 743                                              dev->start_port]);
 744                 break;
 745         default:
 746                 break;
 747         }
 748 }
 749
 750 static void mcast_add_one(struct ib_device *device)
 751 {
 752         struct mcast_device *dev;
 753         struct mcast_port *port;
 754         int i;
 755
 756         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
 757                 return;
 758
 759         dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
 760                       GFP_KERNEL);
 761         if (!dev)
 762                 return;
 763
 764         if (device->node_type == RDMA_NODE_IB_SWITCH)
 765                 dev->start_port = dev->end_port = 0;
 766         else {
 767                 dev->start_port = 1;
 768                 dev->end_port = device->phys_port_cnt;
 769         }
 770
 771         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 772                 port = &dev->port[i];
 773                 port->dev = dev;
 774                 port->port_num = dev->start_port + i;
 775                 spin_lock_init(&port->lock);
 776                 port->table = RB_ROOT;
 777                 init_completion(&port->comp);
 778                 atomic_set(&port->refcount, 1);
 779         }
 780
 781         dev->device = device;
 782         ib_set_client_data(device, &mcast_client, dev);
 783
 784         INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
 785         ib_register_event_handler(&dev->event_handler);
 786 }
 787
 788 static void mcast_remove_one(struct ib_device *device)
 789 {
 790         struct mcast_device *dev;
 791         struct mcast_port *port;
 792         int i;
 793
 794         dev = ib_get_client_data(device, &mcast_client);
 795         if (!dev)
 796                 return;
 797
 798         ib_unregister_event_handler(&dev->event_handler);
 799         flush_workqueue(mcast_wq);
 800
 801         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 802                 port = &dev->port[i];
 803                 deref_port(port);
 804                 wait_for_completion(&port->comp);
 805         }
 806
 807         kfree(dev);
 808 }
 809
 810 int mcast_init(void)
 811 {
 812         int ret;
 813
 814         mcast_wq = create_singlethread_workqueue("ib_mcast");
 815         if (!mcast_wq)
 816                 return -ENOMEM;
 817
 818         ib_sa_register_client(&sa_client);
 819
 820         ret = ib_register_client(&mcast_client);
 821         if (ret)
 822                 goto err;
 823         return 0;
 824
 825 err:
 826         ib_sa_unregister_client(&sa_client);
 827         destroy_workqueue(mcast_wq);
 828         return ret;
 829 }
 830
 831 void mcast_cleanup(void)
 832 {
 833         ib_unregister_client(&mcast_client);
 834         ib_sa_unregister_client(&sa_client);
 835         destroy_workqueue(mcast_wq);
 836 }